beigebox-security

Use when the user wants to deploy / test / restart / debug BeigeBox running in the `beigebox` LXC on the debian host — i.e. "sync source to the LXC", "test in the LXC venv", "restart beigebox", "spawn a long-running task in the LXC", "we have root pane 1", "run pytest in beigebox", "deploy this branch", "what's the LXC's IP". Captures the operational pattern Claude actually uses to drive that LXC from a non-root agent on the debian host. Skip this skill for single-machine / non-LXC tasks; skip for the pop-os Docker BeigeBox.

Agentic task loop. Routes all completions through BeigeBox /v1/chat/completions and all tool calls through /mcp so wiretap captures the full session. Use this to drive multi-step tasks with file access, bash execution, and any registered BeigeBox tool.

Use when the user wants to audit / profile / fingerprint one or more machines to learn OS, hardware, network interfaces, installed virtualization stacks, running containers/VMs, and listening services. Takes a list of targets — local, SSH alias, user@host, or user:password@host — and works against Linux and macOS regardless of authentication style. Produces a markdown report (default), JSON, or a CLAUDE.md-ready section.

Use to load or update accumulated per-host operator knowledge for the BeigeBox project's LAN hosts (pop-os, dssh/debian, assh/mac). Auto-loads at session start when the user references a known host. Auto-reflects at compaction. Notes capture non-obvious, transcript-validated facts (sudo quirks, container IPs, undocumented service contracts) — never secrets, never anything derivable from man/--help.

Use when the user says "make a skill that …", "write a skill for …", "I need a skill that does …", "package this as a skill", or "automate this as a skill" — i.e. wants to turn a recurring task into a new entry under `beigebox/skills/<name>/`. Produces a skill directory in the house style of `services-inventory` and `host-audit`: YAML frontmatter with a `Use when …` description, a `When to invoke` trigger list, a `Usage` block, and (when the skill does deterministic work) a `scripts/` dir with an executable. Run `scripts/scaffold.sh <name>` to stub the layout.

Use when the user says "make a tool that …", "write a tool for …", "I need a tool that does …", "add a tool to BeigeBox", or describes a capability the decision LLM should be able to invoke (search, fetch, lookup, parse, automate). Produces a new tool module under `beigebox/tools/<name>.py` plus the registry hook in `beigebox/tools/registry.py` and a `tools.<name>:` config block in `config.yaml`. Tools are Python classes with a `description` attribute and a single `run(self, input_str: str) -> str` entrypoint, returning user-facing strings (errors included) — they never raise to the caller.

Use when the user asks what services/containers/VMs are running, how to connect to something (URL, IP, shell), or wants an inventory of endpoints across local and remote hosts. Works on Linux and macOS. Probes Docker, Podman, Incus, LXD, classic LXC, libvirt, systemd-nspawn, OrbStack, Colima, Multipass, VirtualBox, Parallels, VMware Fusion, and Tart. Supports SSH to remote hosts and emits either a human table or JSON.

Use when the user wants static analysis / SAST / type-checking / lint-with-teeth / dependency-CVE / secrets scanning on a Python codebase — finding security smells, logic bugs, type errors, vulnerable deps, and leaked credentials without running the code. Wraps ruff (full ruleset including bandit-port S rules), semgrep (registry-backed pattern + dataflow rules), mypy (type checking), pip-audit (SCA / dependency CVE scanning against OSV), and detect-secrets (credential / API-key leaks in source). Emits garlicpress-shape findings so the output merges with fuzz/static pipelines without translation. Five non-overlapping categories: ruff covers AST patterns + bandit; semgrep covers cross-file dataflow; mypy covers types; pip-audit covers known-CVE deps; detect-secrets covers leaked credentials. Skip bandit (ruff -S already covers it).